Let the machines optimize the machines
summary
Claim: Akamas’ machine learning optimizer automatically tunes hundreds of JVM and Kubernetes parameters by iterating apply‑workload‑measure cycles using Locust, Prometheus and Istio. Result: a found configuration improves cost efficiency by 77% in ~24 hours, raises overall throughput ~19% and cuts 90th‑percentile latency ~60%. JVM case: throughput increases 28% before SLO violation after a ~250% heap bump, switch from G1 to Parallel GC and fewer GC threads. The system actually converges and yields specific parameter changes (CPU down for many services, CPU/memory up for catalog and recommender).
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:17System configurations for SREs
- 4:48Performance optimization challenges
- 5:53Automated optimization requirements
- 7:15Automated optimization phases
- 8:40SREs in AI-driven optimization
- 9:00Kubernetes and JVM for Online Boutique
- 14:57JVM tuning for microservices
- 17:26Key takeaways and conclusion
related talks
22mAutomating Performance Tuning with Machine Learning
27mLeverage ML to automatically tune K8s apps
01h 01mContinuous Optimization of MIcroservices using Machine Learning
41mAutomated Performance Tuning with Bayesian Optimization
15mHow can machine learning assist to improve JVM performance?
36mMigrating Applications to the Cloud in a Cost Effective Manner
56mPerformance tuning Twitter services with Graal and Machine Learning
31mFailing to Autoscale?